{"id":"3ecfbc84-32fb-4d4d-8a75-aa33ab55b27b","arxiv_id":"2506.08654","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A federated Pix2Pix model trained on data from three hospitals without sharing images generated synthetic CTs from CBCT on a fourth unseen hospital with MAE 75.22 HU, comparable to training centers.","lead":"Researchers trained a shared AI model on CT data from three hospitals without sharing the images, then tested it on a fourth unseen hospital. The synthetic CTs it produced were about as accurate on the unseen hospital as on the training hospitals, suggesting that privacy-preserving multi-center training is feasible for this task.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Residual misregistration may confound the reported cross-center metrics, so the generalization claim is not yet objectively supported.","rationale":"The reader's weakest-assumption analysis correctly identifies the registration confound as the most load-bearing threat to the central claim. The paper's evidence for generalization is almost entirely quantitative: MAE, SSIM, and PSNR computed on test cases after an additional deformable registration step. The authors explicitly acknowledge that registration errors cannot be objectively quantified and that these errors are the main source of the residual differences they observe, including the lower Center B scores. If registration quality varies by center, then Table 2 ranks centers by alignment quality rather than by model synthesis quality, which would undermine the 'effective generalization across centers' claim. This is not a matter of disagreeing with the community's preferred baselines; it is a direct validity threat to the reported numbers. The proposed test—using an objective registration surrogate and stratifying or regressing on it—would settle the issue. Other limitations, such as the small per-client test sets, the absence of a centralized-training baseline, and the broad privacy claim, are secondary: they affect interpretation of the contribution but do not invalidate the specific quantitative comparison as directly as the unmeasured registration confound does. Because the concern is real but addressable, and because the reader's verdict is already CONDITIONAL, no change in verdict is needed; the condition should explicitly require an objective registration-quality check or a sensitivity analysis that controls for misalignment.","tokens_in":13110,"tokens_out":4313,"duration_ms":59832,"concrete_test":"For every test case in Centers A, B, C, and E, compute an objective registration-quality surrogate after the authors' deformable registration—for example, Dice coefficient between bone segmentations automatically derived from CBCT and CT, or normalized mutual information between the registered volumes. Then (1) regress patient-level masked MAE on this surrogate and (2) recompute Table 2 after restricting the analysis to cases with high registration quality (e.g., the top half by Dice). If the regression is significant or the center-level MAE ordering changes within the well-registered subset, the cross-center generalization claim is confounded by misregistration; if the ordering persists, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the federated model generalized effectively across centers—rests on the image-similarity metrics in Table 2. The paper itself states in Section 4 that CBCT-to-CT misalignment is the primary source of residual error and that no objective method is available to quantify registration errors, leaving only qualitative visual inspection. Because the test data were deformably registered before evaluation, and because head-and-neck images contain high-contrast bone-air interfaces, even small residual misregistration can add large contributions to masked MAE. Center B's lower scores are attributed to a 'higher prevalence of registration issues' and Center C's better scores to 'reduced influence from registration errors,' both based on visual inspection of only four test patients. If registration quality differs systematically across centers, the rank order of Table 2 reflects registration difficulty, not model generalization. The external Center A results are subject to the same confound because they were obtained after the same additional deformable registration step. Without an objective, independent registration-quality measure and without adjusting for it in the analysis, the strongest quantitative evidence for the generalization claim is not established. This is an internal-validity issue rather than a disagreement with external consensus: the reported metrics may be measuring the preprocessing pipeline as much as the federated model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a cross-silo federated learning framework for CBCT-to-synthetic CT (sCT) translation in the head and neck, extending the authors' previous FedSynthCT-Brain work to a new imaging modality. A Pix2Pix conditional GAN is trained on three European centers from the SynthRAD2025 challenge (Centers B, C, E) using FedAvg with FedProx, with Center A held out as an external test set of 60 patients and four test patients per participating center. The reported image similarity metrics (Table 2) show Center A achieving MAE 75.22 ± 11.81 HU, SSIM 0.904 ± 0.034, and PSNR 33.52 ± 2.06 dB, which the authors interpret as evidence of effective cross-center generalization without fine-tuning. The conclusion focuses on the technical feasibility of privacy-preserving federated training for CBCT-to-sCT synthesis.","tokens_in":13299,"tokens_out":5606,"duration_ms":62326,"significance":"If the results are substantiated, this work provides a valuable proof-of-concept that federated learning can support multi-institutional training for CBCT-to-sCT synthesis, a clinically relevant task where data sharing is constrained by privacy regulations. The external validation on 60 patients from an unseen center is a strong design element, and the use of the public SynthRAD2025 dataset supports reproducibility. However, the absence of centralized or single-center baselines, the small test set per participating center (n=4), and the unquantified registration confound currently limit the strength of the generalization claims and prevent the paper from demonstrating an advantage over standard training.","major_comments":[{"comment":"The claim that the federated model 'generalized effectively' is not supported by any comparison with a centralized model trained on pooled data or with single-center models. Without these baselines, the reader cannot determine whether federated training yields a benefit or incurs a performance loss relative to standard training. Please report the same metrics for a centralized baseline and/or single-center baselines using the same architecture, preprocessing, and hyperparameters.","section":"Section 4, Table 2"},{"comment":"Each participating center contributes only four test patients, and the reported standard deviations are large (e.g., Center C MAE = 64.38 ± 13.63 HU). No statistical tests are performed to assess whether the observed differences across centers are meaningful, so the cross-center comparison in Table 2 could be dominated by sampling variability. Please provide per-patient results and appropriate statistical comparisons (e.g., confidence intervals or paired tests against a baseline), or explicitly limit the generalization claim to the external Center A cohort.","section":"Section 3.2 and Section 4"},{"comment":"The paper acknowledges that residual misregistration between CBCT and CT is the primary source of error and attributes Center B's lower scores to 'a higher prevalence of registration issues,' but it also states that no objective method is available to quantify registration errors. Because the test data were deformably registered and the metrics are computed on the registered images, the observed differences among centers could reflect registration quality rather than model generalization. Please provide a quantitative proxy for registration quality (e.g., metrics computed before and after registration, or an independent alignment measure) or substantially temper the cross-center generalization claims.","section":"Section 4"}],"minor_comments":[{"comment":"The text refers to 'Centers B, C, and D' but Center D was excluded in Section 3.1; this should read 'Centers B, C, and E.'","section":"Section 3.5"},{"comment":"The deformable registration algorithm applied to the test data is not described; please specify the method and its parameters so that the evaluation pipeline can be reproduced.","section":"Section 3.2"},{"comment":"The number of federated aggregation rounds is not reported, which makes it impossible to assess convergence or computational cost; please add this information.","section":"Section 3.4"},{"comment":"Equation (3) defines Lpixel as MAE(CT - sCT) but the text calls it Masked MAE; please clarify that the metric is computed within the body mask, consistent with Section 3.5.","section":"Section 3.3"},{"comment":"There is a typo in 'infrastracture'; it should be 'infrastructure.'","section":"Section 3.2"},{"comment":"The statement that 'no objective and independent method is available' to quantify registration errors is overly strong; consider rewording to indicate that no such method was applied in the current study.","section":"Section 5"},{"comment":"The phrase 'comparable performance' on the external validation set is ambiguous; please specify that it is comparable to the participating centers or provide the benchmark explicitly.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable feasibility study with a strong external validation design, but the central claims of 'effective generalization' and competitiveness of federated training are not yet supported because no baselines are included. The registration confound is acknowledged by the authors but is used as an explanation for cross-center differences without quantitative evidence; this is a key internal-validity issue. The small per-center test sets (n=4) further weaken the comparative claims. With the addition of centralized/single-center baselines and a more careful treatment of the registration confound, the paper could become suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the application: federated learning for CBCT-to-sCT synthesis in the head and neck, built on the authors' own FedSynthCT-Brain MRI-to-CT pipeline. As far as I know, it's the first FL result for this task, and they use the public SynthRAD2025 data, which makes the work reproducible. The external validation on 60 patients from Center A, completely held out from training, is a real strength — those metrics are comparable to the participating centers, which supports the feasibility claim.\n\nWhat the paper does well is mostly honesty and clarity. The federation setup is described in enough detail to follow, they use both FedAvg and FedProx with a clear rationale, and they openly discuss the registration problem from the start. The qualitative figures are useful, and the data availability statement points to a public dataset.\n\nThe soft spots are real but not fatal. The biggest one is the registration confound, and the paper essentially concedes it: they say no objective method is available to quantify registration errors, and they attribute Center B's lower scores and Center C's higher scores to unquantified differences in misalignment. That means the rank order in Table 2 may reflect preprocessing difficulty as much as model generalization. The stress-test note is right that this weakens the central claim. The fix is not impossible — they could use a registration-quality surrogate, or at least report metrics on a subset they judge to be well-aligned, but as written the quantitative evidence for cross-center comparability is shaky.\n\nSecond, there are no baselines. No centralized training on all three centers combined, no single-center models, no FedAvg-only comparison. Without those, we don't know whether federation costs accuracy or how much robustness it actually buys. That is a fair demand for a first application.\n\nThird, only four test patients per participating center. The standard deviations are large, and no statistical tests are run, so the per-center differences could easily be noise.\n\nOn the privacy claim: it's standard FL framing, not exaggerated beyond the field's norm, but it would be nice if they acknowledged that sharing model updates and preprocessing parameters still leaks some information.\n\nOverall, this is a useful feasibility study with an honest limitation section. It deserves a serious referee because it breaks new ground, even if the contribution is application-level. I'd send it to review, but the revision should add centralized baselines and a more principled treatment of the registration confound.\n\nMy recommendation: engage with it, but read Table 2 as \"preprocessed-data similarity under federation\" rather than as proof of generalized CBCT-to-sCT quality.","headline":"First FL CBCT-to-sCT feasibility study with a genuinely held-out external test set; the registration confound is acknowledged but under-handled, and missing baselines keep the generalization claim conditional.","tokens_in":13840,"tokens_out":2002,"would_cite":true,"duration_ms":25653,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["87.57.Q-"],"model":"deepseek-v4-flash","headline":"Three hospitals train one CBCT-to-CT model without sharing data","keywords":["federated learning","CBCT-to-sCT synthesis","head and neck radiotherapy","conditional GAN","Pix2Pix","FedAvg-FedProx aggregation","cross-institutional generalization","synthetic CT"],"falsifier":"Compute landmark-based target registration error on the external and low-performing center's test scans and correlate it with per-case MAE. If the low-performing center's disadvantage disappears when only well-aligned cases are compared, the cross-center generalization claim is partly an artifact of registration; if it persists, the federated model itself is the limiting factor.","tokens_in":12893,"feed_emoji":"🏥","tokens_out":6736,"duration_ms":77804,"temperature":0.7,"pith_summary":"This paper sets out to show that federated learning, training a model across hospitals while keeping each site's imaging data local, can produce a CBCT-to-synthetic-CT translator for the head and neck that works across institutions. The authors train a conditional GAN on data from three European centers and evaluate it on a fourth center's 60-patient dataset that never took part in training. The external results (MAE 75.22 ± 11.81 HU, SSIM 0.904 ± 0.034, PSNR 33.52 ± 2.06 dB) fall inside the range of the participating centers, which is what the claim of cross-center generalization rests on. A sympathetic reader would care because radiotherapy dose calculation needs reliable CT-like images, and current deep-learning solutions are usually single-site and cannot pool data under privacy rules.","feed_headline":"Three hospitals train one CBCT-to-CT model without sharing data","feed_subtitle":"External test on 60 patients matched training-center accuracy, pointing toward privacy-preserving multi-hospital radiotherapy.","key_machinery":"The argument is carried by a Pix2Pix-style conditional generative adversarial network: a 2D U-Net generator with eight encoder-decoder blocks and a PatchGAN discriminator, trained with Randomized Multi-2D sampling so each local client sees slices from axial, sagittal, and coronal planes in random order. At inference, the final volume is the median of slice-wise predictions across the three planes. The federation layers on top are server-side FedAvg, which averages client weights in proportion to dataset size, and client-side FedProx, which adds a proximal term (μ = 3) to penalize local drift; a shared preprocessing pipeline, including LUT-based intensity normalization and CT clipping, harmonizes inputs without exchanging images.","core_discovery":"The central discovery the paper argues for is that a single federated conditional GAN can generalize across CBCT scanners and protocols without any site-specific fine-tuning. The federated model produced synthetic CTs with image-similarity metrics on the never-seen external center (MAE 75.22 ± 11.81 HU, SSIM 0.904 ± 0.034, PSNR 33.52 ± 2.06 dB) comparable to those at participating centers (MAE 64.38–85.90 HU, SSIM 0.882–0.922, PSNR 32.86–34.91 dB). The authors attribute most of the remaining error to residual misregistration between CBCT and CT rather than to the model failing to synthesize anatomy. They interpret this as evidence that privacy-preserving collaborative training is technically feasible for CBCT-to-sCT and that a federated model can be applied to a new institution without retraining.","pith_inferences":["A testable extension the paper does not run: ablate the LUT normalization to see whether the harmonized intensity range, rather than federated averaging itself, is what makes the external center match.","Rotating the excluded center, holding out each participating center in turn, would test whether generalization holds for any new site and not only for the external center's protocol mix.","Because registration error is only visually assessed, an independent landmark-based registration metric could separate synthesis accuracy from misalignment and would change how the low-performing center's scores are interpreted.","Clinical adoption would additionally require dose-distribution validation, such as plan dose comparison, which the authors note is impossible with the imaging-only dataset."],"forward_implications":["A hospital can join a collaborative CBCT-to-sCT training effort without uploading patient images, as long as it applies the shared preprocessing locally.","The resulting model can be applied directly to a new center's data, bypassing per-site fine-tuning and the need for center-specific commercial solutions.","Scanner and protocol differences do not by themselves prevent cross-center performance parity, since the external center, which used a different scanner, matched the training centers.","Remaining failures are concentrated in air and dense-bone regions and in misaligned test pairs, so further gains will likely come from targeted training for underrepresented tissues and from better registration evaluation."],"supporting_citations":[{"why":"Provides the multi-center CBCT/CT dataset and scanner/protocol metadata the federation is built on.","marker":"[15]"},{"why":"Establishes the federated synthesis framework and training/inference choices this study extends to CBCT.","marker":"[9]"},{"why":"Supplies the Pix2Pix conditional GAN architecture and the pixel-loss weighting used for the generator.","marker":"[18]"},{"why":"Supplies the FedProx proximal term that stabilizes local updates across heterogeneous centers.","marker":"[32]"},{"why":"Supplies the LUT-based normalization that harmonizes CBCT intensity ranges across clients.","marker":"[31]"},{"why":"Defines the cross-center applicability requirement that the external evaluation is meant to satisfy.","marker":"[8]"}],"fun_headline_variants":["Federated CBCT-to-CT model generalizes without data sharing","One model, three hospitals, no data exchange","Privacy-preserving AI converts CBCT to CT across centers","Federated learning makes synthetic CT work on unseen scanners","Multi-hospital AI for CBCT-to-sCT, no retraining needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation presupposes that residual misalignment between CBCT and ground-truth CT after the added deformable registration is not a systematic, center-specific confound; the paper itself says no objective and independent method is available to quantify registration errors and invokes registration issues to explain the low-performing center's lower scores.","fun_headline_variants_meta":{"raw":{"variants":["Federated CBCT-to-CT model generalizes without data sharing","One model, three hospitals, no data exchange","Privacy-preserving AI converts CBCT to CT across centers","Federated learning makes synthetic CT work on unseen scanners","Multi-hospital AI for CBCT-to-sCT, no retraining needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000596,"raw_usage":{"total_tokens":2873,"prompt_tokens":1109,"completion_tokens":1764,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":725,"completion_tokens_details":{"reasoning_tokens":1680}},"tokens_in":725,"tokens_out":1764,"duration_ms":15054,"temperature":1.0,"reasoning_tokens":1680,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:04:53.575945+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute landmark-based target registration error on the external and low-performing center's test scans and correlate it with per-case MAE. If the low-performing center's disadvantage disappears when only well-aligned cases are compared, the cross-center generalization claim is partly an artifact of registration; if it persists, the federated model itself is the limiting factor.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the multi-center CBCT/CT dataset and scanner/protocol metadata the federation is built on."},{"cited_title":"FedSynthCT-Brain: A federated learning framework for multi-institutional brain mri-to-ct synthesis","cited_arxiv_id":null,"evidence_quote":"Establishes the federated synthesis framework and training/inference choices this study extends to CBCT."},{"cited_title":"Normalization techniques for CNN based analysis of surgical cone beam ct volumes","cited_arxiv_id":null,"evidence_quote":"Supplies the LUT-based normalization that harmonizes CBCT intensity ranges across clients."},{"cited_title":"Huijben, Maarten L","cited_arxiv_id":null,"evidence_quote":"Defines the cross-center applicability requirement that the external evaluation is meant to satisfy."}],"review_version":1}